Comparison of Tissue Classification Performance by Deep Learning and Conventional Methods on Colorectal Histopathological Images

Abstract

The automatic evaluation is essential for the diagnosis and treatment of the disease of pathological images. Computer-aided systems are becoming more common day by day in this area. In this study, multi-class (8 different classes) tissue types were studied in colon cancer histopathological images. Data mining algorithms are used in the diagnosis phase in the health field. As a conventional method, first of all, the properties of the images are extracted and then the texture classification process is performed with data mining algorithms. The Gray Level Co-occurrence Matrix (GLCM), Discrete Cosine Transform (DCT), Local Binary Pattern (LBP) are used in textural feature extraction. Along with these attributes, machine learning algorithms, such as k-nearest neighbors (KNN), support vector machines (SVM), random forests (RF), logistic regression (LR) were used for classification. As another method, to remove the attributes and perform classification at the same time, tissue classification was performed using deep learning (convolutional neural network) on histopathological images. Tissue classification was automated using transfer learning based on ResNet-18 architecture, one of the convolutional neural network architectures. According to the determined feature and classification algorithm, the performance rates are also given comparatively. Our experiments showed that RF classifier with LBP and GLCM features provided 82% accuracy, while the deep learning method based on ResNet-18 architecture achieved 88.5% accuracy.